Intentionality

Intentionality is the aboutness of mental states: the way a thought can be of Vienna, about tomorrow, or directed at a person who no longer exists. Brentano proposed it as the mark of the mental, and Husserl made the analysis of intentional acts the central task of phenomenology, a century before anyone needed to ask whether a machine’s states are about anything.

The modern form of that question was set by Searle’s Chinese Room. A system can manipulate symbols in ways that correspond to meaningful conversation without, Searle argued, its states being about anything; whatever meaning there is derives from the people who interpret it. For language models the question has returned in more tractable forms. Probing studies find internal structure that correlates with the truth of represented sentences. Lederman and Mahowald defend the view that model text has derivative content, much as library text does, while noting that novel text, and especially novel reference, strains the analogy.

I do not try to settle whether these systems have original intentionality, and my arguments about what they can transmit to their users are built not to depend on it. A condition on testimony that turned on intentional states would decide the machine case by stipulation, and would lapse without anything being learned if the systems came to be describable in intentional terms. What I ask instead is third-personal: whether the truth of what a system asserts figures in why it asserts it (the Tracking Minimum). A barometer has no intentional states and meets that condition.

The older question does not go away. The interpretability question I do rely on, whether truth-correlated structure governs what a model asserts or is merely available to be pulled on, is related to it but not the same. And phenomenology’s patient analyses of how meaning is constituted in acts remain a resource that contemporary discussion of machine meaning seldom draws on.